Raise warning for 24 compressed sparse-only models (#1107)

In a recent update, we disabled Cutlass kernels for sparse-only models
vllm-project/vllm#12417. As a result,
sparse-24-only compressed-models are no longer runnable in vLLM.

This PR introduces a warning message to inform users when compression is
enabled in scenarios where sparse-only models are unsupported. This
ensures clarity and avoids unexpected behavior when using sparse-24
configurations with vLLM.

Changes:

- Added a warning to notify users when attempting to enable compression
with sparse-only models in unsupported configurations.

---------

Signed-off-by: Rahul Tuli <[email protected]>

Author: rahul-tuli

src/llmcompressor/transformers/compression/sparsity_config.py

@@ -181,7 +181,11 @@ def is_sparse24_bitmask_supported(
             return False
 
         if not is_model_quantized(model):
-            # non-quantized 2:4 sparse models are supported
+            logger.warning(
+                "Compressed Sparse-only 2:4 models are not supported in vLLM<=0.7.0, "
+                "consider saving with `disable_sparse_compression` set, "
+                "`model.save_pretrained(..., disable_sparse_compression=True)`"
+            )
             return True
 
         # when model is quantized, and has 2:4 sparsity